# Example benchmark config — minimal but complete.
# Copy and edit for real runs. Pre-registration is REQUIRED; the path must
# exist and be committed to git before the run starts.
#
# Run with:
#   uv run python -m tests.benchmarks._framework.cli validate tests/benchmarks/cloudopsbench/configs/cloudopsbench_smoke.yml
#   uv run python -m tests.benchmarks._framework.cli run tests/benchmarks/cloudopsbench/configs/cloudopsbench_smoke.yml --dev
#
# Remove --dev for production runs (enables IntegrityGuard).

benchmark: cloudopsbench

modes:
  - opensre+llm

llms:
  - claude-default  # use whatever opensre is configured for; llm_dispatch.py
                    # will let this be a specific provider once it ships

# Pinned model versions — refused at runtime if a model resolves differently.
# For real runs, replace with real provider snapshots:
#   claude-4-sonnet: claude-sonnet-4-5-20250929
#   gpt-5: gpt-5-2025-08-07
#   deepseek-v3.2: deepseek-chat-v3.2
model_versions:
  claude-default: "(opensre-default)"

# Replication — required ≥3 for stochastic LLMs (Box-Hunter-Hunter Ch 3.4)
runs_per_case: 3

# Parallelism — sized for laptop; bump on AWS
workers: 4

# Hard cost cap — framework aborts cleanly when exceeded
cost_budget_usd: 50.0

# Seeded random case selection — Mechanism 6 (no cherry-picking)
seed: 42

# Where artifacts (report.json + per-case JSON) land
output_dir: .bench-results/example/

# Pre-registration — Phase 0 integrity gate. Path must exist + be non-empty
# AND be committed to git before the run. See:
# ~/DevBox/opensre-notes/opensre-benchmark-framework.md § 0
pre_registration_path: tests/benchmarks/cloudopsbench/configs/preregistrations/cloudopsbench_smoke.yml

filters:
  # Filter to a small slice for first runs. Remove these for the full grid.
  limit: 5
  seen_shape: [true]
  # Leave empty to include all systems / categories
  systems: []
  fault_categories: []

report_formats:
  - json
  - markdown